Anda di halaman 1dari 13

European Journal of Operational Research 247 (2015) 401–413

Contents lists available at ScienceDirect

European Journal of Operational Research

journal homepage:

Discrete Optimization

Scheduling operating theatres: Mixed integer programming vs.

constraint programming
Tao Wang a,∗, Nadine Meskens b, David Duvivier c
University of Lyon, INSA-LYON, University of Jean Monnet Saint-Etienne, DISP - EA 4570, F-69621 Villeurbanne, France
Catholic University of Louvain, Louvain School of Management, B-7000 Mons, Belgium
University of Valenciennes and Hainaut-Cambresis, LAMIH - UMR CNRS 8201, Campus du Mont Houy, F-59313 Valenciennes, France

a r t i c l e i n f o a b s t r a c t

Article history: The daily scheduling of an operating theatre is a highly constrained problem. In addition to standard schedul-
Received 30 January 2014 ing constraints, many additional constraints on human and material resources encountered in real life should
Accepted 5 June 2015
be taken into account. These constraints concern the priority of operations, the affinities between surgical
Available online 22 June 2015
team members, renewable and non-renewable resources, various sizes in the block scheduling strategy, and
Keywords: the surgical team’s preferences/availabilities. We developed two models in our research work, using mixed-
Combinatorial optimization integer and constraint programming respectively. These were compared using a real-life case in order to
Constraints satisfaction determine which one coped better with a highly constrained problem. A cross-comparison of the experi-
Linear programming mental results shows that the mixed-integer programming model provides a better performance using the
OR in health services weighted sum objective function than using the makespan minimization objective function. Conversely, the
Models comparison constraint programming model is better suited to the makespan minimization objective function than to the
weighted sum objective function. The originality of this research lies on three levels: (1) two models are pre-
sented in detail and compared using real data; (2) constraint programming is used to schedule the operating
theatre; (3) some new constraints are taken into account, such as the affinities between team members in
the composition of surgical teams, and the priorities of patients such as diabetics.
© 2015 Elsevier B.V. and Association of European Operational Research Societies (EURO) within the
International Federation of Operational Research Societies (IFORS). All rights reserved.

1. Introduction Healthcare systems are fundamental to European societies, making

a huge contribution to human activity and hence economic devel-
Health systems in many countries are in crisis. This situation is not opment. Health and its preservation are human needs that everyone
new; what is new, however, is the awareness that if the current trend wants to obtain and maintain.
continues most health systems will no longer be viable by 2015 (IBM, During the 1960s and 1970s, the prosperous economic situation in
2006). Belgium is no exception to this trend, even though it is some- Belgium allowed almost unlimited expenditure on health, which led
times said to have one of the best healthcare system in the world. Its to enough people being hired and enough material being bought to
strengths are its almost complete coverage of the population (Durant, ensure the supply of good quality services. However, from the 1980s
2006), the dynamics produced by the global supply system and the to the present day, major rationalizations have been undertaken
actual delivery of care, the quality of health education and healthcare focusing notably on hospitals.
providers, and the increasing professionalization of the healthcare The research reported in this paper focuses on one of the central
management structures (Itinera Institute, 2008). Nevertheless, Bel- and most expensive of hospital activities: the management of op-
gium is losing ground to other countries as regards the overall qual- erating theatres. The operating theatre in a hospital is composed of
ity of its healthcare system. The country was ranked sixth in the Euro operating rooms and one recovery room. This research aims to help
Health Consumer Index for 2013. First place went to the Netherlands, the operating theatre manager improve the organization of these fa-
followed by Switzerland, Iceland, Denmark and Norway (EHCI, 2013). cilities, in particular the assignment of surgical cases. The develop-
ment of an efficient algorithm is required for assigning a set of sur-
gical cases to operating rooms during a period (often a week). This

Corresponding author. Tel.: +33 477448954; fax: +33 477448921.
weekly operating theatre planning and scheduling problem is solved
E-mail addresses: (T. Wang), in two phases. In other words, decisions are made at the tactical (N. Meskens), and then the operational level. First, a planning problem is solved to (D. Duvivier). obtain the date of surgery for each patient, allowing for the
0377-2217/© 2015 Elsevier B.V. and Association of European Operational Research Societies (EURO) within the International Federation of Operational Research Societies (IFORS).
All rights reserved.
402 T. Wang et al. / European Journal of Operational Research 247 (2015) 401–413

availability of operating rooms and surgeons. Second, a daily schedul- research. This is an important constraint for team building that has
ing problem is solved to determine the sequence of operations in each not carried weight in previous research works (except in Meskens
operating room (Fei, Chu, & Meskens, 2009). et al., 2013). It has frequently been noted that the same people work
Managing the operating theatre is a complex task because surgi- together every day in the operating room for long periods of time.
cal cases must be planned and scheduled so as to minimize the costs A strong affinity in teamwork can generate synergy through a co-
of operating rooms and satisfy the needs and requests of surgeons, ordinated effort, and allows each member to maximize his or her
anesthetists and nurses. Satisfying the patient’s needs and managing strengths and minimize his or her weaknesses. Obviously, the surgi-
the material resources must also be taken into account. Moreover, cal team’s effectiveness contributes to better communication and so
human and material resources are in limited supply, and legal regu- to quality and safety in the health care received by patients. A grow-
lations have to be fulfilled. ing number of studies have demonstrated this point (Carney, West,
Our study focuses on the daily scheduling of surgical cases at Neily, Mills, & Bagian, 2010; Kurmann et al., 2010; Sekhar and Manto-
the operational level, taking into account human and material con- vani, 2015; Tibbs & Moss, 2014; Weaver et al., 2010).
straints. The role of human beings in the decision-making process Each resource limitation included in an operating theatre man-
cannot be ignored. The operating theatre employs various teams of agement model leads to an increase in its complexity, in terms of the
workers (surgical, nursing, anesthetic, maintenance, etc.). Given the number of variables and/or the number of constraints and the com-
limitations on resources, scheduling their activities is crucial to the putational time required to solve the optimization problem. In the
efficient performance of their work. Some constraints are linked to case of highly constrained problems, however, where the solutions
restrictions on the availability of resources: these include the open- space might be narrow and fragmentary, exact methods could pro-
ing hours of the operating rooms, the availability of surgeons, anes- vide optimal solution within reasonable time. Constraint program-
thetists and nurses, and the availability and numbers of surgical in- ming (CP) may also be another adequate tool for our highly con-
struments, and recovery beds. Other constraints are linked to the strained target problem. Indeed, the goal of CP is to solve problems
specificities of the resources (e.g. the versatility of the operating when combinatorial optimization has to handle many variables and
rooms, staff qualifications). Similarly, just as the operating rooms may constraints.
be dedicated to certain types of procedures, the medical staff may The major difficulty encountered in solving a combinatorial opti-
also be specialized. The operating theatre manager therefore has to mization problem is the explosion in the number of combinations as
try to harmonize resources and increase the versatility of the staff so the number of variables increases. The main problem when consid-
as to optimize the use of the operating theatre. ering such a huge set of combinations is the amount of computing
The daily scheduling of the operating rooms is a highly con- hours or days required to find an optimal solution. This is not ac-
strained problem. A recent review by Meskens, Duvivier, and ceptable in the context of time-limited decision-making. However,
Hanset (2013) suggests that very few studies of this prob- unlike the classical mathematical approach, the constraint program-
lem have taken into account the constraints of both mate- ming paradigm is based on “reasoning” about the constraints and is
rial and human resources. The only constraints on human re- still efficient when a lot of constraints are involved. This is one of the
sources which are normally considered are the number and reasons why we chose to use constraint programming. This method
availability of surgeons, stretcher-bearers and anesthetists and also has the advantage of providing the user with different modes of
sometimes of nurses and surgeon assistants (Beliën & Demeule- functioning when searching for a solution in the path tree of solutions
meester, 2007; Fei et al., 2009, 2010; Guinet & Chaabane, 2003; (Krzysztof, 2003).
Ghazalbash, Sepehri, Shadpour, & Atighehchian, 2012; Hashemi, Based on logic programming and graph theory, constraint pro-
Rousseau, & Pesant, 2014; Marcon, Kharraja, & Simonnet, 2003; gramming is an alternative to mathematical programming for com-
Roland, Di Martinelly, Riane, & Pochet, 2010; Vijayakumar, Parikh, plex problems that have a slow convergence. It is also an efficient
Scott, Barnes, & Gallimore, 2013; Wang et al., 2014). Recent works approach to solving and optimizing problems in the event of nonlin-
(Marques, Captivo, & Vaz Pato, 2012; Van Huele & Vanhoucke, 2014) ear constraints, logical statements, or non-convex solution space. This
also consider the surgeon’s workload. A very small number of re- includes time-tabling problems, sequencing problems, and alloca-
search papers have taken into account priorities between patients tion or rostering problems (IBM ILOG, 2010). Constraint programming
(Cardoen, Demeulemeester, & Belien, 2009a, 2009b; Min & Yih, has been widely applied in industrial scheduling problems (Baptiste,
2010a, 2010b). In general, the material constraints taken into ac- LePape, & Nuijten, 2001; El Khayat, Langevin, & Riopel, 2006;
count are limited to the number of operating rooms, the availabil- Harjunkoski & Grossmann, 2002; Novas & Henning, 2014), but has
ity of specific materials and the number of beds in the recovery rarely been used in the field of healthcare. Some perspectives in
room (Augusto, Xie, & Perdomo, 2010; Bulgarini, Di Lorenzo, Lori, the literature have recommended the use of constraint program-
Matarrese, Schoen, 2014; Fei, Meskens, & Chu, 2010; Ghazalbash et al., ming for solving nurse rostering problems in operating theatres
2012; Kharraja, Hammami, & Abbou, 2004; Pham, & Kinkert, 2008; (Trilling, 2006), staff scheduling in healthcare (Bourdais, Galinier, &
Santibanez, Begen, & Atkins, 2007; Testi & Tànfani, 2009; Vijayaku- Pesant, 2003), and medical resident scheduling problems (Topaloglu
mar et al., 2013; Van Huele & Vanhoucke, 2014). Stochastic aspects & Ozkarahan, 2011). Recently, Hashemi et al. (2014) as well as Van
have been integrated by some authors in the planning and schedul- Huele and Vanhoucke (2014) used constraint programming based
ing problem of operating theatre such as Denton, Viapiano, and on a column generation approach to resolve a surgery-planning and
Vogl (2007), Hans, Wullink, Van Houdenhoven, and Kazemier (2008), scheduling problem. Zhao and Li (2014) resolved a scheduling of
Min and Yih, 2010a, 2010b), Lamiri (2008), Lamiri, Xie, Dolgui, and elective surgeries problem in an ambulatory surgical center. They
Grimaud (2009). proposed and compared constraint programming and Mixed Integer
However, in real life other constraints are also very important and Nonlinear Programming models to solve the scheduling problem. But
we propose in this research to take them into account. These include the treated problem was not highly constrained. The authors did not
human constraints such as the priority of some surgical cases (for ex- take into account any human or material constraints.
ample, children and people suffering from diabetes should be sched- In order to demonstrate the advantages of applying CP to a highly
uled at the beginning of the day), the preferences of surgeons, nurses constrained problem as against a traditional method, we compare a
and anesthetists, and material constraints such as the versatility of constraint programming model (CP-MOD) to a traditional mathemat-
operating rooms and the availability of medical supplies. ical programming one (MP-MOD).
In particular, the affinity relationship between two persons who In our two models, new constraints have been taken into ac-
work on the same surgical cases has been taken into account in this count, such as the priority of operations, the affinities between
T. Wang et al. / European Journal of Operational Research 247 (2015) 401–413 403

surgical team members, renewable and non-renewable resources, operation-recovery process is a two-stage no-wait flow-shop, so the
various sizes in the block scheduling strategy, and the surgical team’s scheduling of operating theatres should ensure that there will be at
preferences/availabilities. The comparison between the two models least one available recovery bed at the end of each operation. When
is the main contribution. no bed is (or will be) free at the end of the surgery, the operation is
Following this introduction, the problem is defined mathemati- not (or will not be) performed. The patient can be transferred to the
cally in the second part of the paper, which sets out its notations and normal bed only after his/her recovery.
hypothesis. The third and fourth parts describe respectively the MP
and CP scheduling models, and the fifth part compares the two mod- 2.1. Problem statement
els. This comparison is made on the one hand by the minimization
of the makespan, and on the other hand by the minimization of the The notations of input data used for the two models are listed
weighted sum of the completion time. Finally we present our conclu- below:
sions and perspectives.  The set of all the surgical cases

2. The problem O The number of operations: O = |  |

o An operation: o∈
T The set of time slots in a day
In this research, we first need to describe mathematically the t A time slot: t∈T
problem of daily scheduling of surgical cases. The time horizon is set  The set of operating rooms
to one day and divided into T time slots (t ∈{1..T}). In our implementa- R The number of operating rooms: R=|  |
tion each time slot is fixed to 30 minutes, the most common divisor of r An operating room: r∈
S The set of surgeons
the surgery durations that we collected from a field study. This tem-
s A surgeon: s∈S
poral granularity is quite close to reality, but does not substantially A The set of anesthetists
increase the size of the search space. There is a set of R operating a An anesthetist: a∈A
rooms. Each room will then be available for T time slots, which will N The set of nurses
correspond to free periods for the surgeons. n A nurse: n∈N
s The set of operating rooms allocated to surgeon s
In practice, there are two types of operations: elective operations, Ωs The set of operations allocated to surgeon s
which are planned by the surgeon in consultation with the patient; Ωb The set of operations with high priority that start earlier than
and emergency operations, which — as their name suggests — are the others
unplanned and arrive unexpectedly. In general, operations for emer- Ωm The set of operations with medium priorities
Ωe The set of operations with low priority that start later than the
gency patients are performed in dedicated operating rooms; some of
them are allowed to be scheduled together with elective operations, Os The number of operations allocated to surgeon s: Os = | Ωs |
if patients’ clinical conditions are relatively stable. In our model, each Ob The number of operations taking place at the beginning of the
operation included in the set of operations to be assigned is denoted day (high priority): Ob = | Ωb |
by o∈. Om The number of operation with medium priorities : Om = | Ωm |
Oe The number of operations that take place at the end of the day
A set of material and human resources is also required for each (low priority): Oe = | Ωe |
operation. Concerning human resources, the legislation requires the do The duration of operation o (in number of time slots)
presence of at least one surgeon, one anesthetist, and two nurses for ESo The earliest start of operation o (time slots)
each surgical case. The affinity relationships between members of LSo The latest start of operation o (time slots)
Kρ The set of renewable resources
surgical team are taken into account. Some specialized works (Leach,
Kυ The set of non-renewable resources
Myrtle, & Weaver, 2011; Mazzocco et al., 2009; Weaver et al., 2010) k A resource: k∈{ Kρ ∪ Kυ }
showed that affinities in a surgical team had an impact on the perfor- ρ
mok The quantity of the renewable resource k required by operation
mance of the team in terms of communication, cooperation, and co- o
ordination, which itself has a significant impact on the quality of care mvok The quantity of the non-renewable resource k required by
operation o
delivered to the patient. Their estimations are obtained by means of ρ
Mk (t ) The quantity of the renewable resource k, available at moment t
a field survey and interviews with the surgical staff. Since affinity re- Mkv The quantity of the non-renewable resource k, available for the
lationships are not all evaluated equal between members, the lower day
value is thought to be acceptable to qualify a bilateral affinity rela- MS (s, t) The availability of surgeon s at time t: MS (s, t) ∈ {0, 1}
MN (n, t) The availability of nurse n at time t: MN (n, t) ∈ {0, 1}
tionship. Hence an overall evaluation of these bilateral affinity rela-
MA (a, t) The availability of anesthetist a at time t: MA (a, t) ∈ {0, 1}
tionships will contribute to building efficient surgical teams. B The set of recovery beds
We make a distinction between renewable and non-renewable re- b A recovery bed: b∈B
sources. Material resources can be renewable (Kρ ) or non-renewable dbo The duration of the recovery after operation o (in number of
(Kυ ). For example, sterile medical trays, which are a basic require- time slots)
B1,o Beginning of operation o, in an operating room (time slots)
ment for operations, need to be completely sterilized between uses, B2,o Beginning of recovery, in the recovery room, at the end of
and so are non-renewable resources in the course of a day. Human operation o (time slots)
resources are typically renewable resources (Kρ ). C2,o End of recovery of the operation o, in the recovery room(time
We consider the existence of three kinds of priority for opera- slots)
Cmax Makespan, end of recovery, in the recovery room, at the end of
tions: High, Medium and Low. Operations with a high priority (set
the latest operation o (time slots)
Ωb ) start earlier than the other operations; they cover children, di- AffSN(s,n) The affinity relationship between surgeon s and nurse n (∈{0,1})
abetics, ambulatory operations, etc. Operations with a low priority AffSA(s,a) The affinity relationship between surgeon s and anesthetist a
(set Ωe ) are carried out later than the others, and are mainly infec- (∈{0,1})
tious cases that contaminate the room and require more cleaning af- AffNA(n,a) The affinity relationship between nurse n and anesthetist a
ter the operation. We have also added preferential constraints related AffNN(n1 ,n2 ) The affinity relationship between two nurses n1 and n2 (∈{0,1})
to the preferences of the surgeons for particular operating rooms and
to the availability of human resources. As discussed above, three sets of operations were created: Ωb , con-
The limited capacity of recovery beds is taken into account. In taining all the operations that have to be carried out before the other
practice, the ratio of operating rooms to recovery beds is 1–1.5 operations; Ωe , containing those that have to be carried out after
(legal constraint), or 1–2 if precautionary measures are taken. The the other operations; and Ωm , containing those that have no special
404 T. Wang et al. / European Journal of Operational Research 247 (2015) 401–413

Table 1
Example of matrix MS (s, t), giving the availability of Surgeons 1, 2 and 3 between 8.00 and 20.00.

8.00 8.30 9.00 9.30 10.00 10.30 11.00 11.30 12.00 … 19.30

T 1 2 3 4 5 6 7 8 9 … 24
Surgeon 1 0 0 0 0 0 1 1 1 1 … 1
Surgeon 2 1 1 1 1 1 0 0 0 0 … 0
Surgeon 3 0 0 0 1 1 1 1 1 0 … 0

Table 2 rion is linear, it might be possible to use a weighted sum of affinities

Example of affinity matrix Aff(p1 ,p2 ).
 (ki × Aff(p1 , p2 )) between the members in the same team. How-
Anes.1 Anes.2 … Anes.A Nurs.1 Nurs.2 … Nurs.N ever, this leads to the problem of choosing these weights ki . In order
to ensure a minimum affinity between all pairs of the members in
Surg.1 9 8 … 0 9 7 … 8
Surg.2 7 7 … 6 7 7 … 7
each team, we have chosen to transform the affinities into binary val-
… … … … … … … … … ues using a threshold. Thus, the integer values are of no use as the
Surg.S 5 6 … 8 0 9 … 9 input data of our mathematical models since the affinity matrix is
Nurs.1 8 8 … 8 5 … 8
only used to answer a binary choice: whether two persons will work
Nurs.2 3 8 … 9 8 … 7 together in a same surgical team. Directly using integer values can-
… … … … … … … … not enrich the results of the models, but increases the computational
Nurs.N 9 8 … 9 8 3 … burden. If the affinity score is greater than or equal to the threshold,
then the two persons p1 and p2 will be happy to work together. Oth-
erwise, the two persons will refuse to work in a same team. Obviously
a high threshold may hinder team-building due to incompatibility
requirement. In addition, we can define the earliest/latest start times
between members. Therefore, if, for example, the threshold is set to
(ESo and LSo ) for each operation when needed.
5, the affinity matrix described above can be transformed as follows
Operations are defined by a set of characteristics. We assume that
(Table 3). The variables AffSA(s,a), AffSN(s,n), AffNA(n,a), AffNN(n1 ,n2 )
the duration of a particular operation, do , can be predicted in advance.
obtain their values from the transformed matrix. However, the value
Furthermore, for the comfort and safety of the patient, we assume
of this threshold has to be defined. Several approaches are possible.
that the operation starts as soon as possible (ESo ) and no later than
One of them consists in starting from a value that is equal to the aver-
(LSo ) (expressed in time slots).
age value of the team members’ affinities, and transforming the ma-
Renewable resources are available again once the operation ter-
trix into a binary matrix. If there is no solution due to a lack of surgical
minates. For each operation, we know the quantity of renewable re-
ρ teams, then we should try again with a lower threshold value (us-
source required by operation mok , and it will be compared to the
ρ ing, for instance, a dichotomy approach to compute the new value).
amount of available resource Mk (t ) at time t. Unlike renewable re- If there is a solution and there is enough time to restart a resolution,
sources, consumed non-renewable resources cannot be reused by then we can increase the threshold value and try to solve the prob-

other following operations. Only the total use of each resource mvok lem with this value, and so on. The resulting approach is a tradeoff
is compared to the available quantity of the non-renewable resource between a set of pretreatments and mathematical models. All ba-
Mkv for the day. sic tests are implemented in the pretreatment phase to simplify the
The availability of each surgeon is expressed in a matrix, MS (s, t). mathematical modelling.
Table 1 gives an example of such a matrix, and shows that Surgeon
1 is available for operations from 10.30, Surgeon 2 between 8.00 and
10.30 and Surgeon 3 between 9.30 and 12.00. In the same way as for 2.2. Assumptions
the surgeon, two availability matrices are given to describe respec-
tively the availabilities of the anesthetist MA (a, t), and of the nurse A list of assumptions must be enumerated in order to complete
MN (n, t). the description of our framework:
The affinity relationship for teamwork is expressed through an
affinity matrix Aff(p1 ,p2 ), where each row and each column repre- • No surgeon can operate on more than one patient at the same
sents a person (a surgeon, an anesthetist, or a nurse). For each pair time. Similarly, no recovery bed can be occupied by more than
of persons, a score of 0–9 is assigned based on the preferences ex- one patient at the same time.
pressed by them (the lower of two unilateral notes). Mutual incom- • All the scheduled patients are ready for their surgery on the given
patibility is denoted by 0 and strong preference is denoted by 9. day.
Several pretreatments are directly applied to this matrix before its • All scheduled operations have to be performed during the current
use in the mathematical model through a Graphical User Interface. day, i.e. no operation can be postponed.
The full description of all these treatments lies outside of the scope • The induction time for each operation, and the post-operation
of this paper. However, we may mention that their objective is to en- clean-up after the operation, are included in the operating time.
sure the homogeneity of the values to a certain extent, and to ensure • The recovery beds in the recovery rooms are identical; that is, the
that there are sufficient non-null values to avoid any lack of teamwork patient can be transferred to any available recovery bed.
with sufficient affinities. • Once a surgical case has started in an operating room, it cannot be
It might be possible to use the integer values in Table 2 directly interrupted, i.e. there is no pre-emption.
in the mathematical models. Then we need to ensure with appropri- • The time needed at the start of every work day to clean the op-
ate constraints, that a team can be built if the average affinity value erating rooms before all operations is not taken into account. The
between the members is greater than a predefined threshold, which post-operation clean-up time (cleaning time after each operation)
could be the median value5 between 0 and 9 for example. However, is included in the operating time. These ‘setup’ times are indepen-
the average value may “hide” some dissatisfaction between members dent of the operating schedule. However, operations may be se-
in the same team. In another solution, ensuring that the affinity crite- quenced in a particular order, as mentioned above. For example,
T. Wang et al. / European Journal of Operational Research 247 (2015) 401–413 405



Fig. 1. The binary variables OTR(o,t,r).

AffNN(n1 ,n2 )

those which are particularly contaminating, such as iatrogenic in-


fections, will be scheduled at the end of the day.

3. The MP model

The daily scheduling problem of the operating theatre can be

mathematically formulated by a mixed integer programming (MP)
model. The objective is to minimize the makespan of the generated

work schedule.
The operations have to be placed individually; each one is de-

scribed by the number of time-slots for which it will take up a room

over the operating duration. Within a room r, the duration of each op-

eration is given by do , a group of consecutive binary variables show-

Anes. 1

ing the assignment of operation o in this room. Only the binary vari-
ables relative to the operation for the room in question and for the
sequence of time slots occupied by the operation will be set to 1;

the others are all set to 0. In this way, a three-dimensional (O op-



erations, T time slots and R rooms) binary variables matrix is created.

This matrix is represented by OTR(o, t, r). The use of such matrices is

motivated by the fact that one of our objectives is to obtain the de-
tailed schedule of each resource. As illustrated in Fig. 1, binary vari-
Example of affinity matrices AffSA(s,a), AffSN(s,n), AffNA(n,a) and AffNN(n1 ,n2 ) after threshold applied.

ables take the value 1 not only at the beginning of the operation but
throughout its duration in a given room r.

A number of other matrices are described for the members of the



team and the recovery room.

OTR(o,t,r) = 1 if operation o is assigned at time t to operating room r

0 otherwise
OTB(o,t,b) = 1 if operation o is assigned at time t to recovery bed b

0 otherwise


STR(s,t,r) = 1 if surgeon s is assigned at time t to operating room r

0 otherwise
NTR(n,t,r) = 1 if nurse n is assigned at time t to operating room r

0 otherwise



ATR(a,t,r) = 1 if anesthetist a is assigned at time t to operating room r

0 otherwise

The formulation of the model can now be written as:



Minimize Cmax (1a)

C2,o = B2,o +dbo − 1 (1b)



Cmax ≥ C2,o (1c)


 o −1
R LSo +d
do .(do −1)
t.OT R(o, t, r) −



r=1 t=ESo
B1,o = ∀o ∈  (2)

Table 3


B LSo +d
o +dbo −1
dbo .(dbo −1)
t.OT B(o, t, b) − 2
b=1 t=ESo +do
B2,o = ∀o ∈ 
406 T. Wang et al. / European Journal of Operational Research 247 (2015) 401–413

OT R(o, t, r) ≤ 1, ∀r ∈  , ∀t ∈ T (4) 2 × ST R(s, t, r) = NT R(n, t, r), ∀s ∈ S, ∀t ∈ T, ∀r ∈ s (19)
o=1 n=1


OT R(o, t, r) ≤ 1, ∀s ∈ S, ∀t ∈ T (5) AT R(a, t, r) ≤ MA (a, t ), ∀a ∈ A, ∀t ∈ T (20)

r∈s o=1 r=1

+do −1


ST R(s, t, r) = AT R(a, t, r), ∀s ∈ S, ∀t ∈ T, ∀r ∈ s (21)
OT R(o, t, r) = do, ∀o ∈  (6)
r=1 t=ESo

⎢ ⎥ ST R(s, t, r) + NT R(n, t, r) − 1 ≤ A f f SN(s, n)

⎢ t+do−1 ⎥
⎢ OT R(o, τ , r ) ⎥
R T −d
  o +1 ⎢ ⎥ ∀s ∈ S, ∀n ∈ N, ∀t ∈ T, ∀r ∈  (22)
⎢ τ =t ⎥ = 1, ∀o ∈  (7)
⎣ do ⎦
r=1 t=1

ST R(s, t, r) + AT R(a, t, r) − 1 ≤ A f f SA(s, a)

O ∀s ∈ S, ∀a ∈ A, ∀t ∈ T, ∀r ∈  (23)
OT B(o, t, b) ≤ 1, ∀b ∈ B, ∀t ∈ T (8)
NT R(n, t, r) + AT R(a, t, r) − 1 ≤ A f f NA(n, a)
B LSo +d
 o +dbo −1
∀n ∈ N, ∀a ∈ A, ∀t ∈ T, ∀r ∈  (24)
OT B(o, t, b) = dbo, ∀o ∈  (9)
b=1 t=ESo +do

⎢ ⎥ NT R(n1 , t, r) + NT R(n2 , t, r) − 1 ≤ A f f NN(n1 , n2 )

⎢ t+db
o −1 ⎥
⎢ OT B (o, τ , b) ⎥ ∀n1 , n2 ∈ N, n1 = n2 , ∀t ∈ T, ∀r ∈ 
o +1 ⎢ τ =t
B T −db
 ⎥ (25)
⎢ ⎥ = 1, ∀o ∈  (10)
⎣ dbo ⎦
OT R(o, t, r) = 0, ∀o ∈ , ∀r ∈  , ∀t ∈/ [ESo, LSo + do]
b=1 t=1

B1,o + do = B2,o, ∀o ∈  (11) OT R(o, t, r) = 0, ∀s ∈ S, ∀o ∈ s , ∀r ∈/ s , ∀t ∈ T (27)

⎛ LSo2 +do2 −1

 OT B(o, t, b) = 0, ∀o ∈ , ∀b ∈ B, ∀t ∈/ [ESo + do, LSo + do + dbo]
⎜ OT R(o2 , t, r) ⎟
⎜ t=ESo2 ⎟ (28)
B1,o1 ≤ B1,o2 + ⎜1 − ⎟T,
⎝ do2 ⎠ The objective function (1a) and constraints (1b) and (1c) ensure
minimization of the makespan of the operating theatre.
The beginning time of an operation or a recovery is given by con-
∀o1 ∈ b , ∀o2 ∈ m , ∀r ∈  (12)
straints (2) and (3). They allow us to locate the first value 1 of the
variables OTR declared for the operation o in an operating room r or
⎛ LSo3 +do3 −1
⎞ on a recovery bed b. Constraints (4) indicate that two operations can-
 not take place at the same time in the same operating room. Further-
⎜ OT R(o3 , t, r) ⎟
⎜ t=ESo3 ⎟ more, there certainly exists an exact match between every operation
B1,o2 ≤ B1,o3 + ⎜1 − ⎟T, and its surgeon. Thus, the operations allocated to each surgeon, se-
⎝ do3 ⎠
lected from the set of surgeons S, are known in advance. Constraints
(5) prevent any surgeon from conducting two operations at the same
∀o2 ∈ m , ∀o3 ∈ e , ∀r ∈  (13) time in different operating rooms. Constraints (6) and (7) were intro-
duced into the model to express the fact that an operation o has to
take place over do consecutive time slots. Constraints (6) require the

ρ ρ number of variables set to 1 to be equal to do , while constraints (7)
mok OT R(o, t, r) ≤ Mk (t ), ∀t ∈ T, ∀k ∈ K ρ (14)
specify that the variables set to 1 in this interval must be continuous,
r=1 o=1
due to the fact that the integer division of the sum of consecutive 1
by do must be equal to 1. In this way, we ensure that there is only one

OT R(o, t, r) ≤ Mkυ , ∀k ∈ K υ (15) string of consecutive 1, representing the operation (zero everywhere
do else). Constraints (8), (9) and (10) express the conditions at the recov-
r=1 o=1 t=1

 ery room stage, for instance the fact that there is only one patient at a
OT R(o, t, r) ≤ MS (s, t ), ∀s ∈ S, ∀t ∈ T, ∀r ∈ s (16) time in a recovery bed, and that the variables OTB set to 1 have to be
o∈s continuous throughout the interval of time dbo . Constraints (11) en-
sure continuity between two stages of the procedure; the expression

ST R(s, t, r) = OT R(o, t, r), ∀s ∈ S, ∀t ∈ T, ∀r ∈ s (17) on the left indicates the last OTR set to 1 at the operating stage of an
o∈s operation, while that on the right represents the first OTB set to 1 at
the recovery stage of the same operation.

R Constraints (12) and (13) are precedence constraints. Similarly,
NT R(n, t, r) ≤ MN (n, t ), ∀n ∈ N, ∀t ∈ T (18) they express the timing requirements for the operations included
r=1 in the sets Ωb and Ωe . High priority operations should start before
T. Wang et al. / European Journal of Operational Research 247 (2015) 401–413 407

those with medium priorities, but low priority operations after those Earliest start time latest end time
with medium priorities. Constraints (14) and (15) represent the lim-
its on the renewable and non-renewable resources. The difference Operation
lies in the fact that we have to weight these variables by the inverse
of the duration of the operation to obtain comparable values. Con-
straints (16) and (17) express the availability of surgeon s at time t by Start time End time t
MS (s, t). Similarly, constraints (18) and (20) are derived from con-
Fig. 2. An interval variable.
straints (16) and express the availability of nurse and anesthetist at
time t by MN (n, t) and MA (a, t). Constraints (19) and (21) express
the fact that a team consists of a surgeon, an anesthetist and two solve time-tabling and sequencing problems usually characterized by
nurses. Constraints from (22) to (25) describe the implementation of non-linear constraints, logical constraints, and incompatibility con-
affinity relationship in the model. Team members can work together straints. The CP engine is often used as a fast generator of feasible so-
only when all combinations Aff(p1 ,p2 ) are greater than or equal to lutions, but takes a considerable amount of time in terms of finding
the affinity threshold. Constraints (26) and (27) express that one op- optimal solutions. Detailed analysis will be described in the Section 5.
eration performed by surgeon s should be placed in the time win-
dow between its earliest start time and latest end time, only in one 4. The CP model
of the operating rooms allocated to surgeon s. They also ensure that
there is no operation planned when the surgeon cannot operate. Fi- Constraint programming deals not only with logical constraints
nally, constraints (28) ensure that post-operative recovery can only used in constraint satisfaction problems (CSP), but also with mathe-
be placed in the time window between its earliest start time and lat- matical constraints that are usually solved by classical approaches. A
est end time. CP model is defined in three parts. The first part describes all decision
Constraints (7) and (10) should be linearized to be used with Cplex variables; the second gives the domain for each of these variables, in-
solver. One of the solutions is to replace each of the two constraints cluding discrete values; the third contains various sets of constraints,
with three linear constraints and new decision variables. The equiva- representing logical and mathematical relationships between
lent constraints to (7) are designed as following. variables.
x(o, t, r) ≥ OT R(o, t, r) − OT R(o, t − 1, r) ∀o ∈ , ∀t ∈ T (7a)
4.1. Decision variables

x(o, t, r) = 1 ∀o ∈  (7b) The operations are represented by interval variables instead of bi-
t=1 r=1 nary variables in MP-MOD. An interval variable represents an interval
of time during which an operation is performed. In general, it is de-
OT R(o, 0, r) = 0 ∀o ∈ , ∀r ∈  (7c) fined (see Fig. 2) by its inherent attributes, such as earliest start time,
latest end time and duration, but its position in time or its start time
In the same way, the equivalent constraints to (10) are designed are unknowns in CP-MOD. In contrast, an operation’s start time in
as following. MP-MOD should be deduced from output results, once the optimal
 solution is found. The definition of interval variables for operations is
1 if recovery after operation o starts at time t in
y(o, t, b) = bed b given as follows (definition 29).
0 otherwise
INTERVAL operation (o in ) in ESo .. LSo + do − 1 SIZE do
y(o, t, b) ≥ OT B(o, t, b) − OT B(o, t − 1, b) ∀o ∈ , ∀t ∈ T (10a) (29)

B 4.2. Logical relationships
y(o, t, b) = 1 ∀o ∈  (10b)
t=1 b=1 An operation can be performed only when all human and mate-
rial resources are brought together. A six-tuple <o,s,r,n1 ,n2 ,a> is pro-
OT B(o, 0, b) = 0 ∀o ∈ , ∀b ∈ B (10c) posed to represent team building for a given operation. The objective
is to make a preselecting constraint that generates all possible sur-
An overview analysis of MP model is performed to estimate the
gical teams which satisfy affinity constraints. An instance of the six-
size of the scheduling problem. The number of decision variables de-
tuple is composed of an operation o, the surgeon s who is capable
pends on (o+s+n+a)∗t∗r+o∗t∗b, and the number of constraints are
to perform the operation o, an operating room r allocated to s, two
given in Table 4 below. The size of the problem scales up quickly with
nurses n1 , n2 , and an anesthetist a, with all (Aff(p1 ,p2 ) > threshold)
the time-splitting scheme and the number of operations, surgeons,
satisfied. The set of all possible surgical teams can be expressed in
nurses, anesthetists, operating rooms, and recovery beds. A MP opti-
terms of the definition (30). In the same way, the set of all operation-
mization engine will take a considerable amount of time to analyze
recovery combinations is obtained by the definition (31). These two
the model’s feasibility and to calculate the optimal solution in a frag-
definitions also considerably reduce the search space of the schedul-
mented solution space.
ing problem.
 ∀s ∈ S, ∀o ∈ s , ∀r ∈ s , ∀n1 , n2 ∈ N, n1 = n2 , ∀a ∈ A,

Teams = < o, s, r, n1 , n2 , a >  A f f SA(s, a) × A f f SN(s, n1 ) × A f f NA(n1 , a) × A f f SN(s, n2 ) (30)
 ×A f f NA(n2 , a) × A f f NN(n1 , n2 ) > 0

Due to the limits introduced by the use of linearization in solv- Recovery = {< o, b > |∀b ∈ B, ∀o ∈  } (31)
ing scheduling problems, the effectiveness of MP models is appar-
ently not guaranteed, especially for large-scale problems. In con- In one of the solutions generated by the CP-MOD, only one six-
trast, constraint programming offers suitable modeling techniques to tuple per operation will be present. The others, having not satisfied
408 T. Wang et al. / European Journal of Operational Research 247 (2015) 401–413

Table 4
Overview analysis about number of constraints in MP-MOD.

Set of constraints Decision matrix Number of constraints Set of constraints Decision matrix Number of constraints

4 OTR R∗T 17 STR, OTR T ∗ |s |
5 OTR S∗T 18 NTR N∗T
7 OTR O 20 ATR A∗T
10 OTB O 23 STR, ATR S∗T∗R∗A
12 OTR R∗Ob ∗Om 25 NTR N²∗T∗R
13 OTR R∗Om ∗Oe 26 OTR O∗T∗R
14 OTR T∗|Kρ | 27 OTR Os ∗ T ∗ (R − |s |)
15 OTR |Kυ | 28 OTB O∗T∗B

16 OTR T ∗ |s |

all constraints and the objective function, are absent. This can be ex-
pressed by a Boolean function (32).
true, ∀x ∈ Teams, x is present
presenceOf(x) = (32)
false, ∀x ∈ Teams, x is absent 1

Any given resource, no matter whether it is a human resource, an 0 1 2 3

operating room or a recovery bed, can be used at moment t by only
one operation or one patient, so there is no question of overlap be- Fig. 3. Example of stepwise function stepwise{0→1, 1→2, 0→3, 2→4, 1→5}.
tween two operations in the schedule of a resource above. However
an overlap is allowed when two operations use different resources.
Here, all individual schedules of these resources are defined as fol-
h h
lows, but no-overlap constraints will be added in the model’s section
StepAtStart(o,h) StepAtEnd(o,h)
below. Definitions from (33) to (37) give different schedules for sur-
geon s, nurse n1 , n2 , anesthetist a, operating room r, and recovery bed
b as soon as a feasible solution found. Operation o Operation o
Sched(s) = operation(x)∀x ∈ Teams, presenceOf(x), x.s = s
Fig. 4. Example of stepAtStart(o,h) and stepAtEnd(o,h).
∀s ∈ S (33)

Sched (n) = operation (x)∀x ∈ Teams, presenceOf (x),  
Avail(n ∈ N) = stepwise(t ∈ T ) Mn (n, t ) → t; 0 (39)
x.n1 = n or x.n2 = n} ∀n1 , n2 ∈ N (34)

Sched(a) = operation(x)∀x ∈ Teams, presenceOf(x), x.a = a Avail(a ∈ A) = stepwise(t ∈ T ) Ma (a, t ) → t; 0 (40)

∀a ∈ A (35) However, the availabilities of renewable resources cannot be de-

scribed by a simple stepwise function. The quantity of a renewable
   resource is reduced at the beginning of an operation and restored at
Sched(r) = operation(x)∀x ∈ Teams, presenceOf(x), x.r = r
the end of it. Two additional functions are introduced to model the
∀r ∈  (36) consumption and the production of a cumulative resource. The value
of function stepAtStart(o,h) changes to h at the start of an interval
   variable, an operation o in our case, while the value of stepAtEnd(o,h)
Sched(b) = operation(y)∀y ∈ Recovery, presenceOf(y), y.b= b changes at the end of the operation o. Therefore, a function Qty(k) is
∀b ∈ B (37) given in (41) to express the available quantity of renewable resource
k over time (Fig. 4).
The availabilities of each member in a surgical team are expressed
by a stepwise function. A typical stepwise function, denoted by step- Qty(k ∈ K ρ ) = stepwise(t ∈ T ) Mk (t ) → t; 0
wise(i in 1..n) {Value[i] → Time point[i]} is often used to model the  ρ
availability or the use of a resource over time. An example of a step- − stepAtStart(o, mok )+ stepAtEnd(o, mok ) (41)
wise function is illustrated in Fig. 3. Using its concept, the functions o∈ o∈

from (38) to (40) respectively express the availabilities of surgeon,

nurse and anesthetist in time. 4.3. The constraint programming model
Avail(s ∈ S) = stepwise(t ∈ T ) Ms (s, t ) → t; 0
  The CP-Model of the daily scheduling problem is introduced be-
Ms (s, 1), t < 1 low.
= Ms (s, i + 1), ∀i ∈ [1..n − 1], ∀t ∈ [ti , ti+1 ) (38)
0, t > tn MinimizeMax(endOf(o)) ∀o ∈  (42)
T. Wang et al. / European Journal of Operational Research 247 (2015) 401–413 409

s.t. MP model. Furthermore, additional variables and constraints have

been developed for the purpose of obtaining linear constraints. In
¬(start(δ1 ) < start(δ2 ) < end(δ1 )) ⇔ noOverlap(Sched(β))
the case of linear programming, where the solver accepts only linear
∀δ1 , δ2 ∈ Sched(β), ∀β ∈ S ∪ N ∪ A ∪  ∪ B (43) constraint and linear objective function, linearization becomes a ma-
jor obstacle to the modelling work at a conceptual level. There is no
presenceOf(∀x1 ∈ Teams, x1 .o = o) ⇒ ¬(∃x2 ∈ Teams, guarantee that equivalent and efficient linear constraints can always
x2 .o = o, x2 = x1 , presenceOf(x2 )) ∀o ∈  be found. None of these issues matter in the CP model.
Second, the constraint section of the CP model takes into account
only some precedence and overlap constraints; all the others, like
presenceOf(∀y1 ∈ Recovery, y1 .o = o) ⇒ ¬(∃y2 ∈ Recovery, affinity constraints from (22) to (25) in the MP model, can be directly
y2 .o = o, y2 = y1 , presenceOf(y2 )) ∀o ∈  (45) integrated into the definitions of team tuples (30). This can signifi-
cantly reduce the search space of the daily scheduling problem. In
addition, the start time of each operation, defined by (2), becomes
End(operation(x)) + 1 = Start(recovery(y)) an inherent unknown of the interval variables; the availability con-
∀x ∈ Teams, ∀y ∈ Recovery, x.o = y.o (46) straints (16) (18) and (20) can be replaced by (38), (39) and (40) in
the CP model. Constraint programming using logical operators, vari-
ables, and expressions is much more intuitive for the modelling work
End(operation(x1 )) ≥ Start(operation(x2 )) of scheduling problems, compared with mathematical programming.
∀x1 , x2 ∈ Teams, x1 .r = x2 .r,
(x1 .o ∈ b , x2 .o ∈ m ) or (x1 .o ∈ m , x2 .o ∈ e ) 5. Experimental results

Once the two models have been properly formulated and speci-
fied, we are now able to compare them and assess which is the more
successful. The models were both run on a Core2TM Duo processor
presenceOf(x) ⇒ Avail (s) ∧ Avail (n1 ) ∧ Avail (n2 ) ∧ Avail (a) = 1 (2 GHz, 4 GB RAM, Operating System: Windows 7). The MP-MOD,
∀x ∈ Teams, x.s = s, x.n1 = n1 , x.n2 = n2 , x.a = a, using mixed-integer programming, was coded in Optimization Pro-
start(operation(x)) ≤ t ≤ end(operation(x)) gramming Language and solved by Cplex 12.5 (IBM ILOG, 2010).
(48) The CP-MOD uses a constraint-programming method. It was
solved by the Constraint Programming Optimizer included in IBM
ILOG optimizers. This method possesses the descriptive power
Qty(k) ≥ 0 ∀k ∈ K ρ ∪ K υ (49)
needed to model this kind of problem accurately. Built on an event-
The objective function (42) ensures minimization of the makespan based propagation mechanism with back-tracking structure, con-
of the operating theatre. The function endOf(o) gives the end of re- straint programming also offers the advantage of finding a feasible
covery of the operation o, and we have endOf(o) = C2, o . Constraints solution via a depth-first search algorithm, or a more sophisticated,
(43) are used to indicate that no overlap is allowed between two branch-and-bound type of search can be carried out to find an opti-
operations inside an individual schedule of any resource, such as mal solution. However such searches tend to involve extremely long
surgeon, nurse, anesthetist, operating room, and recovery bed. Con- computing time (Fages, 1996).
straints (44) and (45) state that for a given operation, the presence
of one surgical team and one recovery bed is exclusive in a gen- 5.1. Input data
erated solution, hence there does not exist another alternative that
is admitted at the same time. Constraints (46) and (47) are prece- In order to evaluate the proposed methods of improving the prac-
dence constraints. Constraints (46) ensure that recovery starts as tical arrangement of surgical cases in the operating theatre, real life
soon as the operation finishes. Constraints (47) express that opera- data from a Belgian University Hospital are used in this study. In this
tions with higher priority should start before those with lower prior- hospital, there are nine surgical specialties: stomatology, gynecology,
ity. Constraints (48) are used to ensure that a six-tuple can be cho- urology, orthopedic surgery, ENT/otorhinolaryngology, ophthalmol-
sen into a feasible solution only if all necessary human resources are ogy, pediatric surgery, plastic surgery and abdominal surgery. The op-
available during the whole period of operation. Finally, Constraints erating theatre in this hospital is composed of 4 operating rooms and
(49) concern the available quantity of renewable and non-renewable one recovery room with 8 beds. Normally, all the operating rooms
resources. are open from 8 a.m. to 6 p.m., and can be extended to 8 p.m. if nec-
essary. The recovery room opens simultaneously with the operating
4.4. Comparison with the MP model rooms and does not close until the last patient has left the operating
There are various ways to model this daily scheduling problem, The main aim is to provide the operating theatre manager with
and the choice of modeling is crucial to the performance of the model. a good solution, which satisfies all the constraints and can be com-
Our MP and CP models have been optimized but remain accurate to puted in a short time. The data used in this study come from 6321
the real world constraints. The comparison between these two mod- records from the operating theatre, collected over a one-year period.
els is based on our choice of modeling that aims at testing their re- The data consist mainly of date of surgery, induction time, the start
spective abilities. time and end time of surgery, the time the patient left the operating
First, the decision variables are defined in different ways. Binary room, the surgeon and specialty for each surgical case, the reason for
variables are used in the MP model to describe each of the five de- admittance and some personal information (the patient’s birthday,
cision variables in a three-dimensional matrix. The core element of gender, etc.). Overtime hours are not considered because we have de-
this daily scheduling problem, the operations, is represented by a se- fined the time slots for a working day as fixed. However they could be
quence of OTR (o,t,r). However in the CP model, an integrated interval taken into account by allocating them a specific hourly cost, higher
variable is sufficient to describe an operation. than that which applies during the normal day.
In order to ensure that an operation is correctly represented in The same surgical cases undertaken by the same surgeons using
output results, Constraints (6) and (7) should be introduced into the the same resources were analyzed in each model, with the duration
410 T. Wang et al. / European Journal of Operational Research 247 (2015) 401–413

Table 5
Data sets.

Data sets Operations Surgeons Anesthetists Nurses Renewable resources Non-renewable resources Operating rooms Recovery beds

D1 8 4 4 8 5 5 4 8
D2 15 7 4 8 5 5 4 8
D3 17 9 4 8 5 5 4 8

Table 6
Comparison of solutions obtained through MP-MOD.


Data set Affinity threshold Number of constraints Number of variables Number of solutions Makespan Time (opt.)

D1 5 17061 6251 5 14 26.41

6 17061 6251 4 14 26.22
≥7 Infeasible

D2 5 23350 10662 4 24 10.05

6 23350 10662 2 24 30.59
≥7 Infeasible

D3 5 26620 12032 2 24 27.89

6 26620 12032 1 24 32.74
≥7 Infeasible

Table 7
Comparison of solutions obtained through CP-MOD.


Data set Affinity threshold Number of constraints Number of variables Number of solutions Makespan Time (best) End time

D1 5 207929 1900 1 14 8.01 387.36

6 28985 616 1 14 2.52 94.3
≥7 Infeasible

D2 5 715985 3541 1 23 12.24 728.85

6 45591 928 1 23 1.51 115.87
≥7 Infeasible

D3 5 868313 4011 1 23 16.17 772.45

6 49263 1029 1 23 2.18 133.98
≥7 Infeasible

of the time slots set to 30 minutes (the greatest common factor of matrix inside a data set, the numbers of constraints and variables do
operation durations). Besides the performance comparison between not change along with the threshold value for a given data set, but in-
two models, their robustness is also evaluated using three datasets crease from a small data set to a larger one. In D1, eight surgical cases
of various sizes (from D1 to D3 in Table 5), with all constraints taken need to be scheduled; if the threshold value is set to 5 or 6, the objec-
into account. For example, the dataset 3 for this day consists of 17 tive function leads to an optimal solution to 14 time-slots within less
operations, 9 surgeons, 8 nurses, 4 anesthetists, 5 renewable material than 27 seconds. Using D2 and D3, the optimal solution found - also
resources and 5 non-renewable resources. The differences between within a reasonably short time - is that all surgical cases can be fin-
datasets consist in the numbers of operations and of surgeons. ished before the end of 24th time-slot. When the threshold value goes
As mentioned earlier in Section 2.1, the integer affinity matrix is beyond 6, no any feasible solution can be found, because no surgi-
transformed into a binary matrix by using a threshold value. In the cal teams can be formed. The column ‘Number of solutions’ indicates
presented results the threshold value is set to 5, the median value. It that the number of feasible solutions found before the solver termi-
will then be increased successively in a sequence of testing scenarios nates. There are a few solutions which can satisfy all the constraints.
as shown in Table 6. This phenomenon, as mentioned at the beginning of the paper, cor-
responds to the main characteristic of highly constrained problems.
5.2. Computational results with the makespan objective Table 7 shows the results generated by the CP-MOD. With re-
gard to constraints, their number is much larger in CP-MOD than in
The objective function in both models is defined by a simple for- MP-MOD, because the CP engine needs to generate more domain in-
mulation that minimizes the makespan, that is, the maximum job formation about the variables from the compact formulation of the
completion time, as presented by the functions (1a) and (42). Tables problem, in order to perform a filtering algorithm. In consequence,
6 and 7 compare the computational results of the two models. much more memory is required. However, fewer variables are em-
Table 6 gives the results obtained from experiments using the MP- ployed since an operation is represented by only one interval variable,
MOD. All the three data sets have been introduced into the model suc- instead of a sequence of binary variables. An important phenomenon
cessively, and for each of these data sets, different threshold values should be noted; the numbers of constraints and variables decrease
have been applied. Since the threshold value only affects the affinity when the threshold value increases from 5 to 6, as opposed to what
T. Wang et al. / European Journal of Operational Research 247 (2015) 401–413 411

Table 8
Comparison of solutions obtained through MP-MOD.


Data set Affinity threshold Number of constraints Number of variables Number of solutions Makespan Time (opt.)

D1 5 17052 6241 5 14 1.97

6 17052 6241 3 14 1.64
≥7 Infeasible

D2 5 23334 10645 2 23 2.01

6 23334 10645 1 23 1.96
≥7 Infeasible

D3 5 26602 12013 1 23 3.59

6 26602 12013 1 23 3.95
≥7 Infeasible

Table 9
Comparison of solutions obtained through CP-MOD.


Data set Affinity threshold Number of constraints Number of variables Number of solutions Makespan Time (best) End time

D1 5 207929 1900 2 14 7.8 757.69

6 28985 616 2 14 1.9 240.94
≥7 Infeasible

D2 5 715985 3541 1 23 25.16 2229.42

6 45591 928 3 23 19.14 314.66
≥7 Infeasible

D3 5 868313 4011 1 23 44.14 1862.62

6 49263 1029 12 23 133.78 407.98
≥7 Infeasible

we have indicated in Table 6. The problem size is then reduced. As And in the CP-MOD
a result, the solver can find a feasible solution much more rapidly.  
Minimize w1 endO f (o) + w2 endO f (o)
In other words, the CP-MOD is highly sensitive to restrictions, even
o∈Ob o∈Om
when they are configured inside the input data. As for the solutions’

quality, only one solution has been identified in each scenario, but + w3 endO f (o) (51)
this first and best solution was found slightly quicker than the cor- o∈Oe
responding computing time of the MP-MOD. In the cases where the where w1 , w2 , and w3 are non-negative weights and w1 >>w2 >>w3 .
threshold value is set to 6 in D2 and D3, the CP-MOD gives a bet- Tables 8 and 9 compare the computational results after imple-
ter solution than the MP-MOD in less than 3 seconds. The column menting this new objective function.
‘End time’ indicates the time spent before the CP-MOD reaches the Table 8 presents better results compared with those in Table 6.
fail limits1 . CP-MOD is obviously inefficient in identifying optimal so- There is indeed a very small difference in terms of problem size after
lutions, due to long computing time. But it does not seem necessary changing the objective function, while optimal solutions are found
to seek the optimum solution for a highly constrained and frequently ten times faster than before. However, it should be noted that this
encountered problem, like the daily scheduling problem. weighted sum objective function remains a specific function specially
designed for our daily surgical scheduling problem. In contrast, our
initial objective function (minimizing the makespan) is a more gen-
5.3. Computational results with a weighted sum objective function eral function fitting most scheduling problems.
Table 9 compares the results obtained from the CP-MOD after
The makespan objective function used initially in the two models changing the objective function to a weighted sum. With regard to the
is a logical function rather than a mathematical one. Now that the CP- problem size, none of the numbers of constraints and variables has
MOD with the makespan objective seems to be better than the MP- been changed, compared to Table 7. Though the same best makespan
MOD, it becomes necessary to estimate whether the CP-MOD retains can always be found in each experiment, the amount of time spent
an advantage under a pure mathematical objective function. A second has increased. Especially, the ‘End time’ becomes very long. The CP
objective function is then designed as follows. solver takes much more time to examine the feasibility of one poten-
In the MP-MOD tial solution. In addition, when D3 is used and the threshold value
   is set to 6, the CP-MOD returned 12 feasible solutions, among which
Minimize w1 C2,o + w2 C2,o + w3 C2,o (50) 10 refer to the same makespan (i.e. 23). That means, a lot of alterna-
o∈Ob o∈Om o∈Oe tives have been found which satisfy all the constraints including all
the precedence constraints. Only the order of several surgical oper-
ations is different between two alternatives. In real life, any of these
Considering extremely long computing time before finding the optimal solution
alternatives conforms to what we expect in an operating theater; nev-
with CP optimizer, we set the fail limit to 200,000 in the CP-MOD as a stopping rule. ertheless, having more alternatives is conducive to the management
That means 200,000 failures can occur before terminating the search. of an operating theater facing unexpected events.
412 T. Wang et al. / European Journal of Operational Research 247 (2015) 401–413

6. Conclusions and perspectives Bulgarini, N., Di Lorenzo, D., Lori, A., Matarrese, D., & Schoen, F. (2014). Operating room
joint planning and scheduling. In A. Matta, J. Li, E. Sahin, E. Lanzarone, & J. Fowler
(Eds.), Proceedings of the International Conference on Health Care Systems Engineering
In this paper, we introduced two models, using mixed-integer lin- (pp. 127–138). Springer Proceedings in Mathematics & Statistics.
ear programming and constraint programming respectively, to solve Carney, B. T., West, P., Neily, J., Mills, P. D., & Bagian, J. P. (2010). Differences in nurse
the daily surgical case scheduling problem at the operational level, and surgeon perceptions of teamwork: implications for use of a briefing checklist
in the OR. AORN Journal, 91(6), 722–729.
taking into account human and material constraints. In particular, Cardoen, B., Demeulemeester, E., & Belien, J. (2009). Optimizing a multiple objective
affinity relationships between surgical team members have been in- surgical case sequencing problem. International Journal of Production Economics,
tegrated into the problem formulation and modelling, an inclusion 119, 354–366.
Cardoen, B., Demeulemeester, E., & Belien, J. (2009). Sequencing surgical cases in a day-
that makes this scheduling problem a highly constrained problem.
care environment: an exact branch-and-price approach. Computers & Operations
Regarding the detailed formulation of the two models, the CP-MOD Research, 36, 2660–2669.
allows logical constraints that provide explicit formulation and intu- Denton, B., Viapiano, J., & Vogl, A. (2007). Optimization of surgery sequencing and
scheduling decisions under uncertainty. Health Care Management Science, 10, 13–
itive comprehension, while the MP-MOD can only be solved once all
the constraints are linearized. Our MP-MOD, designed for a real-life Durant, G. (2006). Financement des hôpitaux: Des divergences, mais surtout des con-
highly constrained problem, can be solved to optimality in a short vergencesˇ, Revue, N°1 p. 7.
computation time. EHCI (2013).
Using real-life data, three data sets having different sizes have El Khayat, G., Langevin, A., & Riopel, D. (2006). Integrated production and material han-
been designed and used to generate a series of experimental sce- dling scheduling using mathematical programming and constraint programming.
narios, in each of which a different threshold value has been applied European Journal of Operational Research, 175(3), 1818–1832.
Fages, F. (1996). Programmation par contraintes, Ellipses Marketing.
to transform the affinity matrix into a binary matrix. In addition, a Fei, H., Chu, C., & Meskens, N. (2009). Solving a tactical operating room planning prob-
weighted sum objective function which takes into account opera- lem by a column-generation-based heuristic procedure with four criteria. Annals
tions’ priorities has been introduced to carry out an extended analy- of Operations Research, 166(1), 91–108.
Fei, H., Meskens, N., & Chu, C. (2010). A planning and scheduling problem for an oper-
sis. We have made a cross-comparison through experimental results. ating theatre using an open scheduling strategy. Computers and Industrial Engineer-
On the basis of the results, a few feasible solutions have been identi- ing, 58(2), 221–230.
fied for most scenarios, with the MP-MOD giving better performance Ghazalbash, S., Sepehri, M. M., Shadpour, P., & Atighehchian, A. (2012). Operating
room scheduling in teaching hospitals. Advances in Operations Research, 2012, 1–
using the weighted sum objective function than the initial makespan
objective function. In contrast, the CP-MOD works with the initial ob- Guinet, A., & Chaabane, S. (2003). Operating theatre planning. International Journal of
jective function rather than the weighted sum objective function. Production Economics, 85, 69–81.
Hans, E., Wullink, G., Van Houdenhoven, M., & Kazemier, G. (2008). Robust surgery
In general, the mathematical programming model has the advan-
loading. European Journal of Operational Research, 185, 1038–1050.
tage that numerous methods of resolution already exist in terms of Harjunkoski, I., & Grossmann, I. E. (2002). Decomposition techniques for multistage
investigating the solution space. However it has to be composed of scheduling problems using mixed-integer and constraint programming methods.
linear constraints, in order to be solved using MILP solver. More- Computers & Chemical Engineering, 26(11), 1533–1552.
Hashemi, D., Rousseau, L., & Pesant, G. (2014). A constraint programming based column
over, it does not allow the sophisticated constraints encountered in generation approach for operating room planning and scheduling. Lecture Notes in
real life without complicating the model in such a way that it be- Computer Science, 8451, 455–463.
comes difficult to build, explain, generalize and solve. In contrast, IBM (2006). Santé 2015: Gagnant-gagnant ou perdant-perdant. https://www-
the constraint programming model uses in addition a declarative perdant_perdant_sommaire.pdf.
logic language, which has a higher descriptive power than classical IBM ILOG (2010).
mixed-integer linear programming languages. Logical and non-linear Itinera Institute (2008). Quel est l’état de santé du système de soins en Bel-
constraints can be included. It allows a large number of constraints press/hc/.
to be taken into account. In particular, the more constraints there are, Kharraja, S., Hammami, S., & Abbou, R. (2004). Plan directeur d’allocation des plages
the better the CP model will perform, due to the fact that problem size horaires: approche globale. Gestion et ingénierie des systèmes hospitaliers. Mons,
Belgium: GISEH’04 9-11 September.
decreases, as does the computation time required to find feasible so-
Krzysztof, R. A. (2003). Principles of constraint programming. Cambridge: Cambridge
lutions. Certainly, the CP model is usually slower than the MP model University Press.
in finding optimal solutions, but it is more efficient in generating fea- Kurmann, A., Tschan, F., Semmer, N. K., Seelandt, J., Candinas, D., & Beldi, G. (2010).
Human factors in the operating room – The surgeon’s view. Trends in Anaesthesia
sible solutions, especially for highly constrained and frequently en-
and Critical Care, 2(5), 224–227.
countered problems. Lamiri, M. (2008). Planification des blocs opératoires avec prises en compte des aléas.
Future perspectives focus on integrating additional real-life con- France: University of St Etienne, Doctoral thesis.
straints, such as size of surgical team according to specific operations, Lamiri, M., Xie, X. L., Dolgui, A., & Grimaud, F. (2009). A stochastic model for operating
room planning with elective and emergency demand for surgery. European Journal
skill of surgical nurse, nurses’ work schedule, and overtime hours. of Operational Research, 185, 1026–1037.
This will take us further away from “standard/academic” problems Leach, L. S., Myrtle, R. C., & Weaver, F. A. (2011). Surgical Team: role perspectives and
and toward real-life highly constrained problems. Matheuristics are role dynamics in the operating room. Health Services Management Research, 24(2),
surely one of our perspectives if larger instances occur and/or if the Marcon, E., Kharraja, S., & Simonnet, G. (2003). The operating theatre planning by the
number of constraints continues to increase. follow-up of the risk of no realization. International Journal of Production Economics,
85, 83–90.
Marques, I., Captivo, M. E., & Vaz Pato, M. (2012). An integer programming approach
to elective surgery scheduling analysis and comparison based on a real case. OR
spectrum, 34, 407–427.
References Mazzocco, K., Petitti, D. B., Fong, K. T., Bonacum, D., Brookey, J., Graham, S., Lasky, R. E.,
Sexton, J. B., & Thomas, E. J. (2009). Surgical team behavior and patient outcomes.
The American Journal of Surgery, 197(5), 678–685.
Augusto, V., Xie, X., & Perdomo, V. (2010). Operating theatre scheduling with limited Meskens, N., Duvivier, D., & Hanset, A. (2013). Multiobjective scheduling of operat-
recovery beds and patient recovery in operating rooms. Computers & Industrial En- ing theatre considering desiderata of surgical team. Decision Support Systems, May
gineering, 58, 231–238. 2013, 55(2), 650–659.
Baptiste, P., LePape, C., & Nuijten, W. (2001). Constraint-Based scheduling – applying con- Min, D., & Yih, Y. (2010). Scheduling elective surgery under uncertainty and down-
straint programming to scheduling problems. Kluwer Academic Publishers. stream capacity constraints. European Journal of Operational Research, 206, 642–
Beliën, J., & Demeulemeester, E. (2007). A branch-and-price approach for integrating 652.
nurse and surgery scheduling. European Journal Operational Research, 189(3), 652– Min, D., & Yih, Y. (2010). An elective surgery scheduling problem considering patient
668. priority. Computers & Operations Research, 37, 1091–1099.
Bourdais, S., Galinier, Ph., & Pesant, G. (2003). HIBISCUS: a constraint programming Novas, J. M., & Henning, G. P. (2014). Integrated scheduling of resource-constrained
application to staff scheduling in health care. Principles and Practice of Constraint flexible manufacturing systems using constraint programming. Expert Systems with
Programming, LNCS, 2833, 153–167. Applications, April 2014, 41(5), 2286–2299.
T. Wang et al. / European Journal of Operational Research 247 (2015) 401–413 413

Pham, D.-N., & Kinkert, A. (2008). Surgical case scheduling as a generalized job shop Trilling, L. (2006). Aide à la décision pour le dimensionnement et le pilotage de
scheduling problem. European Journal of Operational Research, 185, 1011–1025. ressources humaines mutualisées en milieu hospitalier. Institut National des Sciences
Roland, B., Di Martinelly, C., Riane, F., & Pochet, Y. (2010). Scheduling an operating the- Appliquées de Lyon (INSA), Doctoral thesis.
atre under human resource constraints. Computers & Industrial Engineering, 58(2), Van Huele, C., & Vanhoucke, M. (2014). Analysis of the integration of the physician
212–220. rostering problem and the surgery scheduling problem. Journal of medical systems,
Santibanez, C., Begen, M., & Atkins, D. (2007). Surgical block scheduling in a sys- 28(43), 1–16.
tem of hospitals: an application to resource and wait list management in a Vijayakumar, B., Parikh, P. J., Scott, R., Barnes, A., & Gallimore, J. (2013). A dual bin-
British Columbia health authority. Health Care Management Science, 103, 269– packing approach to scheduling surgical cases at a publicly-funded hospital. Euro-
282. pean Journal of Operational Research, 224(3), 583–591.
Sekhar, L., & Mantovani, A. (2015). Teamwork mentality in neurosurgical teams to im- Wang, Y., Tang, J., Pan, Z., & Yan, C. (2014). Particle swarm optimization-based planning
prove patient safety. World Neurosurgery, 83(1), 41–43. and scheduling for a laminar-flow operating room with downstream resources.
Testi, A., & Tànfani, E. (2009). Tactical and operational decisions for operating room Soft Computing doi: 10.1007/s00500-014-1453-z.
planning: efficiency and welfare implications. Health Care Management Science, Weaver, S. J., Rosen, M. A., Diaz Granados, D., Lazzara, E. H., Lyons, R., Salas, E.,
12(4), 363–373. Knych, S. A., McKeever, M., Adler, L., & Barker, M. (2010). Does teamwork improve
Tibbs, S. M., & Moss, J. (2014). Promoting teamwork and surgical optimization: com- performance in the operating room? A multilevel evaluation. Joint Commission.
bining team stepps with a specialty team protocol. AORN Journal, 100(5), 477–488. Journal on Quality and Patient Safety, 36(3), 133–142.
Topaloglu, S., & Ozkarahan, I. (2011). A constraint programming-based solution Zhao, Z., & Li, X. (2014). Scheduling elective surgeries with sequence-dependent setup
approach for medical resident scheduling problems. Computers & Operations times to multiple operating rooms using constraint programming. Operations
Research, 38(1), 246–255. Research for Health Care, 3, 160–167.